The Economics of AI Support: Credits, Models & Cost Per Resolution

This post is the decoder ring. We'll walk the four pricing models that dominate AI customer support in 2026, do the arithmetic that turns each one into a real cost per resolution, and show where credit-based pricing — the model Macha uses — fits. Macha is an AI agent layer that sits on top of the helpdesk you already run (Zendesk, Freshdesk, Gorgias, Front), so the question for us isn't "what does the helpdesk cost" — it's "what does each resolved ticket cost once an AI agent is doing the work." That number is more controllable than the headline rates suggest, and more variable than any vendor's single dollar figure lets on.

The four ways AI support gets priced

There are really only four billing units in this market, plus the platform fee that usually sits underneath them.

The trap is comparing across columns. A "$0.99 per resolution" rate and a "$2.00 per conversation" rate are not two prices for the same thing — they price different events. As Fin's own pricing breakdown puts it, a $1-per-conversation fee costs more than a $1-per-resolution fee whenever your AI resolves fewer than 100% of conversations, because the per-conversation meter still runs on the 40% that fail and escalate. The only number that lets you compare apples to apples is fully-loaded cost per resolved ticket — and getting there means reading the fine print on every model.

Per-resolution pricing, and the definition that decides your bill

Per-resolution (sometimes "outcome-based") pricing is the headline model of the moment, and the logic is genuinely appealing: you only pay when the AI actually closes an issue. Escalations to a human are, in theory, free.

The catch is the word "resolution." Who defines it, and what counts? Intercom's Fin charges $0.99 per resolution, and Intercom's docs define a resolution as either a confirmed outcome (the customer says the answer helped) or an assumed one — the customer simply stops replying after Fin's last message. That assumed-resolution clause matters: a customer who gives up in frustration and never comes back can look identical to a customer who was genuinely helped. Both bill $0.99.

Verified per-resolution rates as of mid-2026:

(Rates from vendor pages and third-party comparisons gathered June 2026 — treat as approximate and re-check the current published terms before you budget. Sources linked at the end.)

Two things worth flagging. First, Zendesk's AI splits its offering: the base "AI Essentials" tier largely surfaces help-center links, while genuinely conversational resolution needs the Advanced AI Agent add-on on top — so the real per-resolution cost is higher than the sticker. Second, at scale these gaps compound brutally. At 100,000 monthly resolutions, the spread between Fin at $0.99 and Zendesk at $1.50 is roughly $51,000 a month. The headline cents are not rounding errors.

Per-conversation and per-seat: the other two

Per-conversation pricing (Tidio's Lyro agent is the canonical example) bills a flat rate for every conversation the AI handles, resolved or not. It's simpler to forecast if your volume is steady, but you explicitly pay for failure — every escalation still costs you the full conversation fee. The model rewards vendors whose AI resolves less, which is an awkward incentive to be on the wrong side of.

Per-seat is the legacy helpdesk model — a fixed monthly license per human agent — and it's increasingly where AI shows up as a bolt-on rather than the main event. It decouples your bill from AI work entirely, which is fine when AI is marginal and misleading when it isn't.

Underneath all of this sits a benchmark worth keeping in view: a human-handled resolution is far from cheap. Channel benchmarks put email/ticket resolution at roughly $6–$16, live chat at $5–$14, and phone at $17–$25 once re-contacts are included; the much-quoted $1–$4 figure is self-service only, not a human touching the ticket (Unthread's channel benchmarks, The Office Gurus 2026 benchmarks). Blended across channels, most teams land around $3–$6 per human-handled resolution. That's the number every AI pricing model is implicitly competing against. If an AI resolution costs you $0.99 and a blended human one costs ~$5, the AI pays for itself the moment it resolves anything a human would have had to.

Credit-based pricing: paying per action, not per outcome

Macha uses the fourth model — credits, charged per AI action — and it's worth being precise about what that means, because it's a genuinely different philosophy.

A credit is spent when the AI does a unit of work: it's deducted once per complete agent response. The number of credits per response depends on the model you put behind the agent. As of the June 2026 lineup, the costs are:

The broader range across the full model catalog runs roughly 0.5 to 9 credits depending on what you pick, with GPT-5.4 Mini as the default at 1 credit. You choose the model per agent — and even per conversation — so the cost of an interaction is something you control, not something the vendor decides after the fact.

The honest distinction from per-resolution pricing: credits don't track outcomes. Macha bills for automation and orchestration — the actions an agent takes — not for whether a ticket ended "resolved." A deflection, a triage, a routed-to-human handoff, and a fully tool-driven fix all consume credits according to the work done, not the outcome reached. That's a feature for buyers who want to model spend on work rather than trust a vendor's resolution heuristic, and a trade-off for buyers who specifically want outcome-only billing. We think it's the more honest framing — but you should know which one you're choosing.

Deriving cost per resolution from credits

Here's the arithmetic that turns credits into the apples-to-apples number. Cost per resolution on a credit model is:

(responses per resolution) × (credits per response) × (price per credit)

Two of those you control; the third depends on your plan. On Macha's Professional plan ($699/mo for 10,000 credits), the effective rate is $0.07 per credit; on Starter ($299/mo) it's $0.10 per credit. Let's run three realistic resolution shapes on Professional pricing:

Even the expensive case lands under the cheapest per-resolution headline rate, and the simple-deflection case is an order of magnitude below it. That's not a magic trick — it's the consequence of paying for actions on a cheap default model rather than a flat outcome fee. The lever that moves your number most is model choice: dropping a high-volume FAQ agent from GPT-5 to GPT-5.4 Mini halves its per-response credit cost, and the June 2026 price cut (GPT-5 from 3 credits to 2) meant Pro customers got 50% more GPT-5 responses inside the same allowance overnight — a credit that bought one response now buys three for every two it used to.

The catch — and this is the per-conversation critique turned on Macha — is that credits are spent whether or not the agent resolves. A long, meandering conversation that ends in a human handoff still burns a credit per agent turn. The discipline that keeps cost-per-resolution honest is the same discipline that makes agents good: tight instructions, the right model for the job, and a clean handoff so the agent doesn't keep talking after it should have escalated. Cheap actions only translate to a cheap cost per resolution if your resolution rate is healthy.

Seeing and controlling the spend

Credit models live or die on visibility, so a few of the guardrails matter as much as the rate.

The usage card splits your monthly plan allowance (which resets each cycle, with a progress bar) from any persistent top-up balance, so you always know which bucket you're spending from.

Top-up packs cover the months you run hot. They're one-time purchases (the medium pack is 5,000 credits for $449), they never expire, and they're spent automatically only after your monthly allowance is gone — so a busy month doesn't strand you, and unused top-up credits roll forward instead of evaporating.

99/mo (3,000 credits), Professional at $699/mo (10,000 credits), and a custom Enterprise tier." loading="lazy" style="display:block;width:100%;height:auto;border-radius:10px;border:1px solid #e5e7eb;box-shadow:0 10px 30px rgba(17,24,39,0.10);">99), Professional ($699), and Enterprise — each bundling a monthly credit allowance." loading="lazy" style="display:block;width:100%;height:auto;border-radius:10px;border:1px solid #e5e7eb;box-shadow:0 10px 30px rgba(17,24,39,0.10);">99/mo (3,000 credits), Professional at $699/mo (10,000 credits), and a custom Enterprise tier.99), Professional ($699), and Enterprise — each bundling a monthly credit allowance.

Credit alerts fire by email at 50%, 80%, and 90% of your monthly allowance — once per threshold, not once per admin — so the bill never surprises you. And for batch jobs like Studies — running an AI analysis across thousands of tickets at once — Macha shows an exact pre-run cost estimate — record count, credits per record, and total — and charges only for records that actually process.

For the most predictable budgets, Enterprise plans bypass credit checks entirely and run on a flat custom contract — the right shape for teams that want a fixed line item instead of a metered one.

The cost layer everyone forgets: it's a layer, not a replacement

Here's the watch-out that no single headline rate captures, and the one most specific to Macha: total cost of ownership has more than one layer. A per-resolution vendor's $0.99 is just their slice. You still pay for the platform underneath it.

Macha is explicit about this because it's structural to what we are: Macha is an AI agent layer on top of your existing helpdesk, not a replacement for it. If you run Macha on Zendesk, your true cost per resolution includes your Zendesk seats plus Macha's credits — two line items, not one. That sounds like a disadvantage until you compare it to the alternative: ripping out a helpdesk your team knows to chase a marginally lower headline AI rate is almost never worth it, and the per-seat helpdesk cost is one you're already paying. The layered model means you add AI economics on top of infrastructure you keep, rather than betting the whole stack on one vendor's resolution definition.

What it looks like annualized

Headline cents are easy to dismiss until you run them at volume. Take a team resolving 5,000 tickets a month with AI, and price the AI layer alone (everyone here still sits on top of, or beside, a helpdesk you also pay for). For Macha we assume a blended ~3 credits per resolution on the default GPT-5.4 Mini model — multi-turn conversations plus the unavoidable credits spent on tickets that escalate — at the Professional rate of $0.07 per credit.

(Vendor rates verified June 2026; treat as approximate. Macha's figure is a model, not a quote — escalation-heavy queues or bigger models push credits-per-resolution up, and it excludes the helpdesk seats you keep paying. The point isn't a precise winner; it's that paying for cheap actions on a tuned model lands an order of magnitude below a flat per-outcome fee at scale.)

The breadth point matters as much as the dollars. A per-resolution vendor like Fin prices one event well; Macha prices the full range of agent work — deflections, triage, routing, and tool-driven fixes — across whichever helpdesk you already run, so the annualized number above is the AI layer for your whole queue, not just the slice that ends in a clean resolution.

So when you build the comparison, build it in full:

• Platform / helpdesk subscription (per-seat, you likely already pay this)
• AI agent cost (per-resolution, per-conversation, or credits)
• Add-ons (Zendesk's Advanced AI Agent, voice surcharges, etc.)
• The escalation tax (what you pay on the conversations the AI doesn't resolve)

Only the sum of those four is a real cost per resolution. A vendor quoting you one layer and calling it the price is the most common way these comparisons go wrong.

When credit-based pricing is the wrong fit

To be straight about it: credits aren't the right model for everyone.

• If you want pure outcome-based billing — pay nothing unless a ticket is resolved — a per-resolution vendor matches that intent more directly. Credits bill for work, not results.
• If your volume is tiny and spiky, a metered model can feel like death by a thousand cuts; a flat per-seat add-on may be simpler to reason about.
• If you can't or won't tune agents, credit costs drift upward — an over-powered model on a high-volume agent, or agents that ramble past the point of handoff, quietly inflate your per-resolution number. Credit pricing rewards operators who tune; it mildly punishes set-and-forget.

Credit-based pricing pays off when you value control and granularity — choosing the model per workload, seeing spend per action, capping batch jobs before they run — over the simplicity of a single outcome fee. For most teams running real volume on top of an existing helpdesk, that control is what keeps cost per resolution genuinely low. But it's a choice, and it's worth making with eyes open.

FAQ

What's the difference between per-resolution and credit-based pricing? Per-resolution charges a fixed fee each time the AI closes an issue (e.g., $0.99). Credit-based charges per AI action — a credit is deducted once per agent response, scaled by the model you choose. Per-resolution ties cost to outcomes; credits tie cost to work done, which you derive into a cost per resolution yourself.

How do I calculate Macha's cost per resolution? Multiply responses-per-resolution × credits-per-response × price-per-credit. On Professional ($0.07 per credit), a one-response FAQ deflection on the default GPT-5.4 Mini model costs about $0.07; a three-turn fix costs about $0.21. See the pricing page for current plans and the docs for the credit details.

Do I pay for conversations the AI can't resolve? On a credit model, yes — credits are spent per action whether or not the conversation resolves, so an escalated ticket still consumes credits for the turns the agent took. Tight instructions and clean handoffs keep that cost contained.

Does Macha replace my helpdesk, so I stop paying Zendesk? No. Macha is an AI agent layer on top of Zendesk, Freshdesk, Gorgias, or Front. Your true cost per resolution includes both your helpdesk seats and Macha's credits — but you keep the helpdesk your team already runs.

What happens if I run out of monthly credits? You buy a one-time top-up pack; top-up credits never expire and are spent automatically after your monthly allowance. Email alerts at 50%, 80%, and 90% warn you before you get there, and Enterprise plans bypass credit checks entirely.

The bottom line

There is no single "price of AI support" — there are four pricing models measuring four different events, and the only number that compares them is fully-loaded cost per resolution. Per-resolution rates are clean headlines hiding a definition you don't control. Credit-based pricing hands you the levers — model choice, action-level visibility, pre-run estimates — and asks you to do the arithmetic, which, done honestly, usually lands lower. Either way, do the full TCO: platform, AI, add-ons, and the escalation tax.

Want to run the numbers on your own queue? Start a 7-day free trial, no credit card required, connect your helpdesk, and watch the credit meter on real tickets — or read the billing docs for the per-model breakdown.

Written by Abbas (Customer Support & AI, Macha) · Reviewed by Ankeet Guha (Co-founder & CTO) · Published 2026-06-24 · Last updated 2026-06-24.

Sources: Fin — per-resolution vs per-conversation · Fin — AI agent pricing comparison · Intercom — Fin AI Agent outcomes · Lorikeet — per-resolution platforms · Unthread — cost per resolution channel benchmarks · The Office Gurus — 2026 support cost benchmarks · Intercom — pricing & seats. Vendor rates gathered June 2026; verify current published terms before budgeting.